MIUI10 Heavyweight Release! Kingsoft Kie provides ultra-clear artifact

Recently, Millet held the "history of the most blockbuster new QI debut" conference. At the meeting, including 8, MIUI10 and so on, including a series of new official debut. Among them, the "domestic conscience system," said MIUI10 is particularly interesting, as the first new generation of debut mobile phone system, MIUI10 "full screen" as the core design, for full screen operation to support and optimize all-round, while the full application of AI, Not only can help users easily handle complex operations, but also to bring users a better quality of HD experience. Next, we will explore the HD quality "artifact" behind MIUI10.

KIE: "Artifact" to make the picture clear

This "artifact" that will make the picture clear is Kie (Kingsoft image &ai enhancement). Kie is Kingsoft's recently launched AI-based image enhancement product that can be used in mobile devices or image transmissions to deliver an ultra-clear image visual experience or save up to 3/4 picture transmission bandwidth costs.

Introduction of MIUI official website for Kie

With the increasing resolution of mobile devices, especially mobile phones, the need for higher resolution images is increasing, because low-resolution images, if placed on a high-resolution screen, can make the visual effect worse. If a 720P picture is placed on a 1080P screen, this problem will occur.

Low-quality images under high-resolution screens can make visual effects worse

The traditional solution-interpolation algorithm, including bilinear interpolation and proximity interpolation, but this method often leads to edge aliasing, blurring and so on, will greatly reduce the visual effect of the picture. Kingsoft Kie is based on artificial intelligence technology, the image reconstruction to make its high-definition magnification, suitable for a more high-definition screen, while the Mosaic, Burr area to repair, enhance the picture visual effect.

Comparison of picture effects using Kie image enhancement features

Comparison of picture effects using Kie image enhancement features

Kie's secret weapon: hyper-score and repair

How does Kie do this? Originally, it has two kinds of secret weapon--kie contains super-resolution (hereinafter referred to as "super-score") and repair two techniques, on the Deep convolutional neural network, Kie's super-fractal model constructs a seven-layer network, which can effectively improve the picture quality, remove the image super-time blur, texture is not clear and so on. The deep convolution neural network can reconstruct the details according to the picture, and realize the effect of far exceeding the original image. Kie has done a lot of exploration and experiment in network structure and model training, and achieved excellent image super-score and denoising effect.

Comparison of picture effects using Kie image enhancement features

Detail comparison

Kie three times times picture magnification effect comparison

First, on the network structure, Kie uses mobile net V2 to construct an efficient network model. Kie combined with residual learning ideas, effectively reduce the learning difficulty, the use of cross-layer dense connection structure, conducive to gradient propagation, so that the depth of the network in a shorter period of time to learn a better model. By using pyramid structure, the characteristics of the bottom layer and the top level can be fused more effectively. In addition, in the training loss, the model uses SSIM, the image content loss, and so on, lets the picture produce the sharper detail.

Kie uses a pyramid structure to make the underlying features and top-level features more effectively combined.

Second, in the model training, Kie collected 300,000 high-definition images, to ensure rich content, using a variety of methods to simulate the generation of real-world network images, so that the model in the inference to achieve better results, can adapt to most of the network pictures, to ensure the comprehensive learning. At the same time the data mix, the different size multiples of the image mixed together to support different multiples of the image of HD.

Kie Residual learning structure: speed up training and improve image quality

For the three-channel YUV image, the human eye is the most sensitive to the Y channel, Kie for Y-channel and UV channel respectively using different networks, where Y-channel complexity is higher than the complexity of the UV channel, in order to ensure the quality of human eyes, speed up the training and inference speed.

Kie for, today's headlines and other mainstream network pictures gathered to investigate, found a lot of smaller resolution of the picture, if the current mainstream mobile terminal to match, need to zoom three times times, which means that the picture needs more details to deal with, so more difficult. Kie with a trained 3X network to process the image, through the convolutional network learning and enhancement, and ultimately improve the quality of the picture.

Kie's application Scenario

At present, Kie is integrated into various applications in the form of SDK, including all kinds of handset manufacturer system, super app, browser and so on. Under the same subjective quality, the Kie has the features of small SDK volume, fast inference speed and low memory footprint, and can adapt to many platforms.

In the bottom-level optimization, Kie uses model compression and quantization techniques to effectively reduce the size and inference speed of the SDK. And, using the Kingsoft AI team to build an efficient deep neural network implementation, on the CPU inference speed reached Google Tensorlite two to three times times, memory consumption also greatly reduced, for the Kie Technology promotion laid the foundation.

At the same time, Kie has a powerful cross-platform feature that currently supports Android, IOS, PC platforms, and web-based interfaces with simple and fast access. Can be processed in real-time on the terminal, with XXX fully decoupled, the picture on the mobile terminal can reach the millisecond response, comprehensively improve the user's browsing experience.

Noun Explanation:

MobileNetV2: is an improvement to the MobileNetV1, is a lightweight convolutional neural network.

Ssim:ssim (Structural Similarity index), structural similarity, is a measure of two image similarity index. The indicator was first presented by the University of Texas at Austin's image and video Engineering Laboratory (Laboratory for image and videos Engineering). In the two images used by the Ssim, one is an uncompressed, undistorted image and the other one is a distorted image.

Google tensorlite: TensorFlow lite,google called Lite version TensorFlow is an extended version of TensorFlow Mobile. Despite being a lightweight version, it's still a big step in deploying deep learning on smartphones and embedded devices. Previously, the model embedded deployment on the phone was supported by TensorFlow Mobile Api,tensorflow. TensorFlow Lite should be considered an upgraded version of TensorFlow Mobile.

MIUI10 Heavyweight Release! Kingsoft Kie provides ultra-clear artifact